Langfuse: Reduce Sentry trace duplication

[vprashrex]

Is your feature request related to a problem?
After upgrading to Langfuse v4, every LLM request results in two traces in Sentry: a bare http.server trace and a separate full waterfall trace for the Langfuse LLM generation. This duplication causes confusion and inflates the span volume in Sentry.

Describe the solution you'd like

• Provide Langfuse with a dedicated, isolated TracerProvider to prevent LangfuseSpanProcessor from appearing on the Sentry provider.
• Add a shared, lazily-initialized isolated TracerProvider to the Langfuse constructor at three locations.
• Ensure multi-tenant safety by allowing each project's LangfuseSpanProcessor to filter spans by public_key.
• Keep flush unaffected by using explicit flushing via .flush().

Result in Sentry: a single clean trace per request with no Langfuse span data leaking.